Download Report

EnviroInfo 2013: Environmental Informatics and Renewable Energies
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
Potential and Problems of the Cellular System Approach for
Environmental Modeling and Simulation
Jochen Wittmann
HTW Berlin, University of Applied Sciences, Dept. Environmental Informatics,
Wilhelminenhofstraße 75A, 12459 Berlin, Germany,
e-mail: wittmann@htw-berlin.de
Abstract
By the growing use of geographic information systems (GIS) the use of cellular approaches for modelling and simulation spatial and especially geographical processes and systems gains more and more
importance. This paper gives an overview over the basic definitions and emphasizes a proper differentiation between the very tight definition of real cellular automata and the very universal one for cellular
systems.
On the base of these definitions typical problems are discussed, which raise for modelling spatial objects given by raster or vector data and by representing spatial relationships between these objects.
Furthermore the difference in semantics between neighbourhood relation in GIS and neighbourhood
relations needed for modelling purposes are elaborated.
1.
Cellular Automata and Cellular Systems
For a long number of years cellular automata serve as a paradigm for modelling real world systems. In
the late 60th J.H.Conway developed the automaton “Life” to model processes such as cell nascence and
death caused by aggregation and isolation. Later on a theoretical foundation for the model class of cellular
automata has been developed (e.g. Banks [1] and Wolfram [5]). For the application area of environmental
modelling these approaches recently re-gained actuality. Because of their acquaintance to the raster data
offered by geographic information systems (GIS) a large number of models dealing with geographical
processes are developed on the base of cellular automata. A representative selection of examples for this
type of models is found in the anthology edited by Goodchild [2].
However, within these examples and most of the application models found elsewhere in literature, the
term cellular automaton is used for a very broad range of modelling approaches which suggests an engagement with the goal of more precise differentiations and definitions.
But the following definitions do not only serve theoretical purposes. For the practitioner the mathematical formalism this paper gives in the following section opens the range of successively more detailed
model specification between the rigid and close definition of cellular automata themselves and the very
general definition of a so-called cellular system.
On the one hand, giving these definitions the modeller will obtain a feeling for potential problem fields
in transferring real system dynamics to cellular systems and on the other hand he gets a guideline how to
use the mathematical concepts to model system behaviour more detailed in a formally proper way.
Doing so, the next section gives two definitions, which will be our base for a discussion of practical
modelling problems later on.
1.1 Definitions
A classical definition for a cellular automaton is given by Wunsch in [6]:
A cellular automaton is a 5-tupel
Z ,Z , R m , f ,
with:
1. Z is a set of states, with elements coming from the state alphabet Z.
N , with N as the set of natural numbers.
2. Z is a non-empty set with Z
3. R m is the spatial extension of the automaton with R m
G the set of whole numbers.
4.
f is called the state transition function with f : R m
G m , m=1,2,3, ... and
Z
Z
5. The neighbourhood relation lists for every space element a set of space elements which are
neighbours to the given one:
( - power set):
: Rm
Rm
.
This definition describes an equidistant raster on the (geographical) space Rm with equal cells. In each
of them works a state automaton with an individual internal state and a state transition function. All the
cells work with the identical state transition function. The states of the cells change discrete in time and
value going from state k to state k+1. The states of the neighbours of a cell may influence the state transition.
To apply this concept for real world problems in environmental modelling the following problems will
arise:
1. The discrete alphabet given by the definition opposes to the need of real value variables for system
quantities.
2. Each cell can hold only one state variable. A typical geographic space unit is characterized by a set of
parameters, however.
3. The automaton only models a sequence of states without any further relation to the time axis.
4. The cells do not have any outputs and inputs in contrast to input- and output functions given by system theoretical concepts.
Caused by these restrictions a more general definition for a so-called cellular system is given, which meets
all the requirements of spatial modelling. Again we define in accordance to Wunsch ([6]):
A cellular system is a 6-tupel
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
T , Z , Rm , , f ,
with:
1. time base T is a non empty set with the relation
defining an order on its elements.
2. Z is a freely definable set called the state set for the system.
3.
R m is the spatial extension with R m
G m , m=1,2,3, ... and
G being the set of whole numbers.
4. f is the state transition function with:
f : Rm
Z
T T
Z
It transfers a state z at a space element r
the same state element at time t (t t ).
R m and at a point in time t to a state z of
6. The neighbourhood relation lists for every space element a set of space elements which are
neighbours to the given one:
( - power set):
: Rm
Rm .
In comparison to the cellular automaton we notice the following extensions:
By the additional time set and its ordering relation time-related processes as needed for system
simulation can be mapped.
All the model variables used in the definition are interpretable as vectors. Doing so, the restriction
of only one state variable per cell is revoked.
The variables are defined without any relevant restriction and therefore can be used for universal
modelling purposes.
The neighborhood relation is unrestricted as well.
There is no advice concerning the spatial collocation of the cells.
From these observations we can state that this definition gives really a non-restrictive formal frame for
modelling. On the other hand, however, it gives only few constructive advise how to build a (spatial)
model. The modeller gets all freedom to fill the frame individually, but just this range of freedom leads to
conceptual problems which shall be discussed in the following sections.
2.
Basic problems in modelling with cellular systems
By practical reasons, the semantics of space used in modelling environmental systems is mostly prespecified by the concepts given by geographic information systems. The simple cause is that all the data to
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
use within the model will come from there. So we first will have a look on the means to model spatial systems and relationships given by recently used GIS.
Geographic information systems provide two very different basic concepts for describing spatial objects: raster data and vector data.
2.1 Modelling with raster data
In the case of raster data a free choosable but equidistant raster is put over a geographic region. It is obvious to identify each of these raster cells with one of the cells of a cellular automaton.
So the advantages of this representation are perspicuously:
easy to handle
GIS data are easily transformed into model data
Free scalable model accuracy by free scalable raster width.
In addition to these arguments the approach is easily adaptable to higher dimensions of space by going
from squares to cubes, from cubes to hypercubes and so on.
Problems using raster data are:
The raster has to be equidistant and perpendicular and the width of the raster determines the scale of the
model.
In the first case the problem lies in the effect of discretisation if an object lies cross to the raster. Caused
by this argument the modeller will be forced to work with a quite small raster width to reach sufficient accuracy.
But doing so, another problem arises: There will be some real world objects to model that spatially extend
over a set of cells. Figure 1 gives an impression of the situation for the modeller between raster representation and vectorized objects.
Fig. 1: values of model quantities in raster and vector representation
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
For those real world objects, which extend over a set of raster cells, a transformation function is needed:
Firstly to summarize the values for this spatial object from the attribute values of the cells covered by the
object. This has to be done by a separate aggregation function which is not mentioned in the standarddefinition of a spatial system.
The analogous problem has to be solved for the other direction: An attribute value of a spatial object has
to be mapped on the set of cells the spatial objects consists of. Not in all the cases a simple homogenous
distribution models the real world circumstances correctly.
A typical example for these problems are population distribution over a geographical and administrative
region such as a municipality: To work with statistical data the extension of the spatial unit has to be the
community itself, the distribution of the people within this region will not be homogenous at all and will
be one of the main interests of the model.
Let us transform this situation to raster GIS in which the municipality is represented by a set of raster
cells:
1. First direction, the aggregation function: The over-all population is the sum of the population of
all the cells.
2. Second direction, the distribution function: How does the population distribute over the cells representing the municipality?
Both functions have to be modelled in addition to the definitions of cellular systems and/or cellular automata!
As these deliberations show, in this modelling task lies the complete scaling problem which is very well
known from the geographic modelling part. The aggregation functions and the distribution functions are
not trivial at all (for a more detailed discussion of this range of problems see Ortmann in [3] or Wittmann
in [4]) and lead to hard complications in contrary to the seemingly very simple modelling approach by cellular systems.
2.2 Modelling with vector data
The alternative is given by a vector-representation of geographical objects as used for vector-GIS. It is obvious that the discretisation problems will completely disappear using this approach. The modelling accuracy is determined by the accuracy of the defining vector line. An aggregation function and/or a distribution function between the different scales of spatial object and raster cells is no longer necessary.
However, this modelling approach suffers under the free editable border line of a spatial model object.
Without any regularity given by a raster, these borders are freely editable now. The consequences are:
1. Regions of different sizes:
The values for all the variables used in the model have to be normalized before brought into relation to values of any other spatial model object.
2. Very complicated neighbourhood relations because of the irregular size of the objects: The number
of neighbours of a cell will be variable which causes a much more complicated specification in the
model description part. The necessary information strongly exceeds the simple definition of cellular automata.
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
This first argument requires the attention of the modeller but can be solved with reasonable effort. The
second argument shall be discussed in detail by the following section.
2.3 Neighbourhoods
For cellular automata and for the corresponding raster-GIS representation the selection of a suited
neighbourhood relation is normally restricted to a selection among the classical relations (von Neumann,
Moore, …). It should be remarked here that the modeller has much more freedom in modelling concerning
the neighbourhood relation, however. Any relation that can be expressed by a filter or a window given by
a set of cell indices in relation to the index of the cell under observation is allowed by the definition. This
can be done very simply by the regularity of the raster.
For cellular systems and vector GIS the neighbourhood specification will change substantially: There is a
completely irregular mosaic over the geographical space given by free definable border lines of the spatial
objects. The neighbourhood is no longer defined by indices of raster cells but by an individual set of
neighbours for each cell.
A typical formal (and mathematical) representation of the neighbourhood relation for all the spatial objects
used within a model would be a neighbourhood matrix as shown exemplarily in figure 2.
Fig. 2: An example for a neighbourhood matrix for a spatial area with six sub-regions
Beneath these implementational remarks using vector representation substantial conceptual problems will
arise at two points:
Geographical versus logical neighbourhood
All neighbourhood relation given by cellular automata has to be interpreted as a geographical neighbourhood. This might result in some discrepancies for modelling. One could imagine that two cells are near
neighbours concerning a Moore-neighbourhood but concerning the semantical relation needed for the
model they are completely isolated to each other.
To give an example: two rural regions are spatial neighbours but separated by a motorway. Therefore the
logical neighbourhood which models reachability is completely different to the one given by the geographical position of the regions under observation. Because of the regularity of the cellular automaton,
this situation cannot be modelled by this concept. Only one neighbourhood relation is allowed. And this
neighbourhood is given by the spatial adjacency of the cells.
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
Using cellular systems this situation could be represented by an individual mark in the neighbourhood matrix. The matrix could be set by the logical neighbourhood relation needed for the modelling purpose.
Again this concept is open for any interpretation (logical neighbours) but requires additional information
(specification of the semantical neighbourhood) on the other hand.
Differentiating logical neighbourhoods
The second problem deals with the differentiation of the neighbourhood relation for the different components of the state vector of a cell in space. Because the vectorization is only allowed for cellular systems,
this problem will appear for this more extended concept only.
To give an example for the problem: A model for the settlement of a population over a geographical region will refer to the attributes “shopping facilities” and “leisure facilities” for a given cell. The facilities
are determined in relation to the existence of shopping centres and recreational facilities in the neighbourhood of the cell under observation. It is obvious that the inhabitant of a cell feels the shopping neighbourhood completely different to the leisure neighbourhood: People will accept a quite long car ride to practice
leisure activities but they like the comfort buying their bread rolls in direct neighbourhood to their residence. This differentiation cannot be expressed by cellular automata at all.
The mightiness of cellular systems would allow a mathematical representation of this situation: For each
of the attributes given in the state vector a differentiated neighbourhood relation can be given. The
mathematical apparatus allows this by specifying a vector of neighbourhood matrixes.
For practical reasons the expenditure in defining all these neighbourhoods has to be taken into account,
however. Although GIS functionality can support this task technically, much deliberation is necessary for
a reasonable specification of the attribute-dependent neighbourhood relationship between the cells. This
should a modeller take into account considering the cost –value ratio for his model.
It should be pointed out that there is a fundamental difference between modelling in GIS and systems
modelling with dynamical processes. GIS administers the geographical relations of spatial objects. Simulation models extend this specification to logical attributes and relations, which should be stored separately. Under this point of view the geographical neighbourhood represents just one of the open set of
“logical” or “semantical” neighbourhoods needed for system modelling.
3.
Conclusion
For the practitioner in modelling there are two general ways to build a model with geographical related
objects using cellular approaches:
First would be the “easy-to-do-approach” by cellular automata. This formal concept gives a very strict
frame for the modelling task and can easily be handled algorithmically. However, for most of the applications its formal expressiveness does not fulfil the needs of the modeller.
So for real world modelling projects more or less extensions are necessary. These extensions can be summarized under the formal concept of cellular systems. All restrictions are revoked by this definition. The
modelling task however gets much more complicated by the design decisions the open concept of cellular
systems demands. Taking into account the deliberations about the representation of spatial objects and the
neighbourhood relations the modeller has to decide the following two points of interest:
1. Homogeneous cells versus individually sized cells.
2. Homogeneous neighbourhood for all the components of the state vector versus attribute specific
“logical” neighbourhoods for each state variable representing the situation of a cell.
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5
In dependence on the decision between the free specifiable cellular system and the restrictive cellular
automaton the expense for implementation of the simulation algorithm and for setting a consistent initial
state for the model has to be considered.
This paper has the intention to give an overview over the range of modelling spatial systems using cellular
formalisms. Main intention was
to give criteria for proper use of the terms cellular automaton in contrary to cellular system and
to give hints for a developer of such models by showing on the one hand the options in modelling and
on the other hand the difficulties the respective decisions will cause.
To give an impression about the dangers when GIS data structures are used for modelling purposes
without thinking about the consequences on semantics in model specification.
Especially for modelling and simulation environmental models the concepts given by cellular systems
form a proper formal and mathematical frame, which should be used to approach spatial modelling problems. However, nothing is as easy as it seems in the beginning: Even using the easy understandable concept of cellular systems, the necessary extensions for modelling real world problems lead to complex and
highly sophisticated model specifications!
4.
Literature
[1]
R.Banks: Cellular Automata, MIT, 1970
[2]
M.F.Goodchild at al. (Ed.): GIS and Environmental Modeling : Progress and Research Issues; GIS
World Books, Fort Collins, Colo, 1996.
[3]
J.Ortmann: Ein Konzept zur individuenorientierten Modellierung von Populationsdynamiken in
Ökosystemen. In: A.Kuhn, S.Wenzel (Hrsg.): Simulationstechnik: 11.Symposium in Dortmund
1997, Vieweg, Braunschweig/Wiesbaden, 1997, S.327-332.
[4]
J.Wittmann: Ein Diskussionsbeitrag zur Problematik der Validierung von individuenorientierten
Modellen. In: H.B.Keller, R.Grützner, M.Sonnenschein (Hrsg.): Werkzeuge für Simulation und Modellbildung in Umweltanwendungen, Forschungszentrum Karlsruhe, 1997.
[5]
S.Wolfram: Theory and Application of Cellular Automata, World Scientific, Singapore, 1986.
[6]
G.Wunsch: Zellulare Systeme, Akademieverlag Berlin, 1977
Copyright 2013 Shaker Verlag, Aachen, ISBN: 978-3-8440-1676-5